- Home
- Archives
- 2026
- Vol. 6 Issue 6
- Article
Machine Learning in Sports: A Question-Centred Review of Methods for Outcome Prediction, Performance Evaluation, Tactics, Athlete Health and Video Analysis
Authors
1 2 Assistant Physical Director, Sethu Bhaskara Agricultural College and Research Foundation, Karaikudi, Tamilnadu, India. 3 Professor and Head, Department of Physical Education and Health Sciences, Alagappa University, Karaikudi, Tamilnadu, India. 4 Assistant Professor, Department of Computational Studies, Kristu Jayanti (Deemed to be university), Bengaluru, Karnataka, India.
Article Information
- Published In
- Volume 6, Issue 6 (2026)
- Published Online
- November-December 2026
- Pages
- 181-188
- Article Type
- Review Article
Abstract
View PDFSport now produces data on almost every action. Leagues record match events, tracking systems log player positions many times a second, wearable sensors measure training load and heart rate, and broadcast video captures every game. Clubs, coaches and analysts use machine learning to turn these data into decisions. The problem is that the literature is divided by sport and by data type, so it is hard to see which methods answer which questions, how well they work, and where they fail. Reviews of match prediction rarely discuss injury risk, and reviews of wearables rarely discuss tactics. This review organizes the field around five questions that analysts ask: who will win, how good is a player, what should the team do, is the athlete fit, and what happened on the field. For each question it compares the machine learning methods used, the data they need and the results they report. Match outcome prediction accuracy varies widely by sport: reported best accuracies reach 86% in American football and 91.8% in ice hockey, but only 59.2% for three-way soccer results and 61.7% balanced accuracy in cricket. In tennis, eleven models ranged from 59% to 72% accuracy, and an Elo model reached 70%. Player valuation has moved from goals and assists to the value of every action. An AI assistant's corner-kick tactics were preferred by experts in 90% of cases. Injury prediction results range from an area under the curve of 0.52 to 0.87, and tree-based ensembles are the most common method. Common weaknesses are small datasets, inconsistent evaluation, limited interpretability and the privacy of athlete data. Future work should build open benchmarks for each question, report calibrated and interpretable predictions, and train on athlete data under privacy protection.
Related Articles
Share Article
*Instagram doesn't support direct link sharing from web. Copy the link and share it in your Instagram story or post.